Fix division by zero when running the tutorial experiment

[d-led]

Problem

when running

% python benchmarks/run.py -e randrot_noise_10distinctobj_surf_agent
# or
% python benchmarks/run_parallel.py -e randrot_noise_10distinctobj_surf_agent

the following run-time errors can be observed:

.../tbp/monty/frameworks/models/motor_policies.py:1259: RuntimeWarning: divide by zero encountered in scalar divide
  return -np.degrees(np.arctan(x / z))
.../numpy/core/fromnumeric.py:3464: RuntimeWarning: Mean of empty slice.
  return _methods._mean(a, axis=axis, dtype=dtype,
.../numpy/core/_methods.py:192: RuntimeWarning: invalid value encountered in scalar divide
  ret = ret.dtype.type(ret / rcount)

as these might affect numeric results, an analysis is necessary. This has implications for published benchmarks that were affected.

Analysis

>>> import numpy as np
>>> x=-1
>>> z=0
>>> -np.degrees(np.arctan(x / z))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ZeroDivisionError: division by zero

interestingly in the shell context this is a run-time error. Is this somehow (mis-)configured to be ignored?

Fix

Fix orienting_angle_from_normal to not be affected by division by zero

Result

> python benchmarks/run.py -e randrot_noise_10distinctobj_surf_agent
...
 overall/percent_correct 100

[tristanls]

FYI: @nielsleadholm, @vkakerbeck, @hlee9212, @scottcanoe

[vkakerbeck]

Wow, thanks for taking a look at this and finding this fix!
We will validate this against some of our longer benchmark experiments.

• If it all looks good and it indeed improves the benchmark results we will open a corresponding PR to update the results table (since currently we don't have a way for external people to run on comparable infrastructure to have the timing column make sense).
• If it doesn't affect performance it will be a nice fix anyways and we can just merge without updating the results table.
• If it impacts performance negatively we'll let you know.

(If you just want to get an impression yourself of how accuracy is affected on more challenging experiments you could also try running one of these benchmarks yourself and just ignore the runtimes when comparing https://thousandbrainsproject.readme.io/docs/benchmark-experiments#results-1)

[d-led]

> python benchmarks/run.py -e base_77obj_dist_agent
...
overall/percent_correct 95.671

vs 93.51% in the table. Eager to see complete results